Thu, 20 Sep 2012 16:49:17 +0200
7023898: Intrinsify AtomicLongFieldUpdater.getAndIncrement()
Summary: use shorter instruction sequences for atomic add and atomic exchange when possible.
Reviewed-by: kvn, jrose
1 /*
2 * Copyright (c) 1997, 2012, Oracle and/or its affiliates. All rights reserved.
3 * DO NOT ALTER OR REMOVE COPYRIGHT NOTICES OR THIS FILE HEADER.
4 *
5 * This code is free software; you can redistribute it and/or modify it
6 * under the terms of the GNU General Public License version 2 only, as
7 * published by the Free Software Foundation.
8 *
9 * This code is distributed in the hope that it will be useful, but WITHOUT
10 * ANY WARRANTY; without even the implied warranty of MERCHANTABILITY or
11 * FITNESS FOR A PARTICULAR PURPOSE. See the GNU General Public License
12 * version 2 for more details (a copy is included in the LICENSE file that
13 * accompanied this code).
14 *
15 * You should have received a copy of the GNU General Public License version
16 * 2 along with this work; if not, write to the Free Software Foundation,
17 * Inc., 51 Franklin St, Fifth Floor, Boston, MA 02110-1301 USA.
18 *
19 * Please contact Oracle, 500 Oracle Parkway, Redwood Shores, CA 94065 USA
20 * or visit www.oracle.com if you need additional information or have any
21 * questions.
22 *
23 */
25 #ifndef SHARE_VM_OPTO_MATCHER_HPP
26 #define SHARE_VM_OPTO_MATCHER_HPP
28 #include "libadt/vectset.hpp"
29 #include "memory/resourceArea.hpp"
30 #include "opto/node.hpp"
31 #include "opto/phaseX.hpp"
32 #include "opto/regmask.hpp"
34 class Compile;
35 class Node;
36 class MachNode;
37 class MachTypeNode;
38 class MachOper;
40 //---------------------------Matcher-------------------------------------------
41 class Matcher : public PhaseTransform {
42 friend class VMStructs;
43 // Private arena of State objects
44 ResourceArea _states_arena;
46 VectorSet _visited; // Visit bits
48 // Used to control the Label pass
49 VectorSet _shared; // Shared Ideal Node
50 VectorSet _dontcare; // Nothing the matcher cares about
52 // Private methods which perform the actual matching and reduction
53 // Walks the label tree, generating machine nodes
54 MachNode *ReduceInst( State *s, int rule, Node *&mem);
55 void ReduceInst_Chain_Rule( State *s, int rule, Node *&mem, MachNode *mach);
56 uint ReduceInst_Interior(State *s, int rule, Node *&mem, MachNode *mach, uint num_opnds);
57 void ReduceOper( State *s, int newrule, Node *&mem, MachNode *mach );
59 // If this node already matched using "rule", return the MachNode for it.
60 MachNode* find_shared_node(Node* n, uint rule);
62 // Convert a dense opcode number to an expanded rule number
63 const int *_reduceOp;
64 const int *_leftOp;
65 const int *_rightOp;
67 // Map dense opcode number to info on when rule is swallowed constant.
68 const bool *_swallowed;
70 // Map dense rule number to determine if this is an instruction chain rule
71 const uint _begin_inst_chain_rule;
72 const uint _end_inst_chain_rule;
74 // We want to clone constants and possible CmpI-variants.
75 // If we do not clone CmpI, then we can have many instances of
76 // condition codes alive at once. This is OK on some chips and
77 // bad on others. Hence the machine-dependent table lookup.
78 const char *_must_clone;
80 // Find shared Nodes, or Nodes that otherwise are Matcher roots
81 void find_shared( Node *n );
83 // Debug and profile information for nodes in old space:
84 GrowableArray<Node_Notes*>* _old_node_note_array;
86 // Node labeling iterator for instruction selection
87 Node *Label_Root( const Node *n, State *svec, Node *control, const Node *mem );
89 Node *transform( Node *dummy );
91 Node_List &_proj_list; // For Machine nodes killing many values
93 Node_Array _shared_nodes;
95 debug_only(Node_Array _old2new_map;) // Map roots of ideal-trees to machine-roots
96 debug_only(Node_Array _new2old_map;) // Maps machine nodes back to ideal
98 // Accessors for the inherited field PhaseTransform::_nodes:
99 void grow_new_node_array(uint idx_limit) {
100 _nodes.map(idx_limit-1, NULL);
101 }
102 bool has_new_node(const Node* n) const {
103 return _nodes.at(n->_idx) != NULL;
104 }
105 Node* new_node(const Node* n) const {
106 assert(has_new_node(n), "set before get");
107 return _nodes.at(n->_idx);
108 }
109 void set_new_node(const Node* n, Node *nn) {
110 assert(!has_new_node(n), "set only once");
111 _nodes.map(n->_idx, nn);
112 }
114 #ifdef ASSERT
115 // Make sure only new nodes are reachable from this node
116 void verify_new_nodes_only(Node* root);
118 Node* _mem_node; // Ideal memory node consumed by mach node
119 #endif
121 // Mach node for ConP #NULL
122 MachNode* _mach_null;
124 public:
125 int LabelRootDepth;
126 // Convert ideal machine register to a register mask for spill-loads
127 static const RegMask *idealreg2regmask[];
128 RegMask *idealreg2spillmask [_last_machine_leaf];
129 RegMask *idealreg2debugmask [_last_machine_leaf];
130 RegMask *idealreg2mhdebugmask[_last_machine_leaf];
131 void init_spill_mask( Node *ret );
132 // Convert machine register number to register mask
133 static uint mreg2regmask_max;
134 static RegMask mreg2regmask[];
135 static RegMask STACK_ONLY_mask;
137 MachNode* mach_null() const { return _mach_null; }
139 bool is_shared( Node *n ) { return _shared.test(n->_idx) != 0; }
140 void set_shared( Node *n ) { _shared.set(n->_idx); }
141 bool is_visited( Node *n ) { return _visited.test(n->_idx) != 0; }
142 void set_visited( Node *n ) { _visited.set(n->_idx); }
143 bool is_dontcare( Node *n ) { return _dontcare.test(n->_idx) != 0; }
144 void set_dontcare( Node *n ) { _dontcare.set(n->_idx); }
146 // Mode bit to tell DFA and expand rules whether we are running after
147 // (or during) register selection. Usually, the matcher runs before,
148 // but it will also get called to generate post-allocation spill code.
149 // In this situation, it is a deadly error to attempt to allocate more
150 // temporary registers.
151 bool _allocation_started;
153 // Machine register names
154 static const char *regName[];
155 // Machine register encodings
156 static const unsigned char _regEncode[];
157 // Machine Node names
158 const char **_ruleName;
159 // Rules that are cheaper to rematerialize than to spill
160 static const uint _begin_rematerialize;
161 static const uint _end_rematerialize;
163 // An array of chars, from 0 to _last_Mach_Reg.
164 // No Save = 'N' (for register windows)
165 // Save on Entry = 'E'
166 // Save on Call = 'C'
167 // Always Save = 'A' (same as SOE + SOC)
168 const char *_register_save_policy;
169 const char *_c_reg_save_policy;
170 // Convert a machine register to a machine register type, so-as to
171 // properly match spill code.
172 const int *_register_save_type;
173 // Maps from machine register to boolean; true if machine register can
174 // be holding a call argument in some signature.
175 static bool can_be_java_arg( int reg );
176 // Maps from machine register to boolean; true if machine register holds
177 // a spillable argument.
178 static bool is_spillable_arg( int reg );
180 // List of IfFalse or IfTrue Nodes that indicate a taken null test.
181 // List is valid in the post-matching space.
182 Node_List _null_check_tests;
183 void collect_null_checks( Node *proj, Node *orig_proj );
184 void validate_null_checks( );
186 Matcher( Node_List &proj_list );
188 // Select instructions for entire method
189 void match( );
190 // Helper for match
191 OptoReg::Name warp_incoming_stk_arg( VMReg reg );
193 // Transform, then walk. Does implicit DCE while walking.
194 // Name changed from "transform" to avoid it being virtual.
195 Node *xform( Node *old_space_node, int Nodes );
197 // Match a single Ideal Node - turn it into a 1-Node tree; Label & Reduce.
198 MachNode *match_tree( const Node *n );
199 MachNode *match_sfpt( SafePointNode *sfpt );
200 // Helper for match_sfpt
201 OptoReg::Name warp_outgoing_stk_arg( VMReg reg, OptoReg::Name begin_out_arg_area, OptoReg::Name &out_arg_limit_per_call );
203 // Initialize first stack mask and related masks.
204 void init_first_stack_mask();
206 // If we should save-on-entry this register
207 bool is_save_on_entry( int reg );
209 // Fixup the save-on-entry registers
210 void Fixup_Save_On_Entry( );
212 // --- Frame handling ---
214 // Register number of the stack slot corresponding to the incoming SP.
215 // Per the Big Picture in the AD file, it is:
216 // SharedInfo::stack0 + locks + in_preserve_stack_slots + pad2.
217 OptoReg::Name _old_SP;
219 // Register number of the stack slot corresponding to the highest incoming
220 // argument on the stack. Per the Big Picture in the AD file, it is:
221 // _old_SP + out_preserve_stack_slots + incoming argument size.
222 OptoReg::Name _in_arg_limit;
224 // Register number of the stack slot corresponding to the new SP.
225 // Per the Big Picture in the AD file, it is:
226 // _in_arg_limit + pad0
227 OptoReg::Name _new_SP;
229 // Register number of the stack slot corresponding to the highest outgoing
230 // argument on the stack. Per the Big Picture in the AD file, it is:
231 // _new_SP + max outgoing arguments of all calls
232 OptoReg::Name _out_arg_limit;
234 OptoRegPair *_parm_regs; // Array of machine registers per argument
235 RegMask *_calling_convention_mask; // Array of RegMasks per argument
237 // Does matcher have a match rule for this ideal node?
238 static const bool has_match_rule(int opcode);
239 static const bool _hasMatchRule[_last_opcode];
241 // Does matcher have a match rule for this ideal node and is the
242 // predicate (if there is one) true?
243 // NOTE: If this function is used more commonly in the future, ADLC
244 // should generate this one.
245 static const bool match_rule_supported(int opcode);
247 // Used to determine if we have fast l2f conversion
248 // USII has it, USIII doesn't
249 static const bool convL2FSupported(void);
251 // Vector width in bytes
252 static const int vector_width_in_bytes(BasicType bt);
254 // Limits on vector size (number of elements).
255 static const int max_vector_size(const BasicType bt);
256 static const int min_vector_size(const BasicType bt);
257 static const bool vector_size_supported(const BasicType bt, int size) {
258 return (Matcher::max_vector_size(bt) >= size &&
259 Matcher::min_vector_size(bt) <= size);
260 }
262 // Vector ideal reg
263 static const int vector_ideal_reg(int len);
265 // CPU supports misaligned vectors store/load.
266 static const bool misaligned_vectors_ok();
268 // Used to determine a "low complexity" 64-bit constant. (Zero is simple.)
269 // The standard of comparison is one (StoreL ConL) vs. two (StoreI ConI).
270 // Depends on the details of 64-bit constant generation on the CPU.
271 static const bool isSimpleConstant64(jlong con);
273 // These calls are all generated by the ADLC
275 // TRUE - grows up, FALSE - grows down (Intel)
276 virtual bool stack_direction() const;
278 // Java-Java calling convention
279 // (what you use when Java calls Java)
281 // Alignment of stack in bytes, standard Intel word alignment is 4.
282 // Sparc probably wants at least double-word (8).
283 static uint stack_alignment_in_bytes();
284 // Alignment of stack, measured in stack slots.
285 // The size of stack slots is defined by VMRegImpl::stack_slot_size.
286 static uint stack_alignment_in_slots() {
287 return stack_alignment_in_bytes() / (VMRegImpl::stack_slot_size);
288 }
290 // Array mapping arguments to registers. Argument 0 is usually the 'this'
291 // pointer. Registers can include stack-slots and regular registers.
292 static void calling_convention( BasicType *, VMRegPair *, uint len, bool is_outgoing );
294 // Convert a sig into a calling convention register layout
295 // and find interesting things about it.
296 static OptoReg::Name find_receiver( bool is_outgoing );
297 // Return address register. On Intel it is a stack-slot. On PowerPC
298 // it is the Link register. On Sparc it is r31?
299 virtual OptoReg::Name return_addr() const;
300 RegMask _return_addr_mask;
301 // Return value register. On Intel it is EAX. On Sparc i0/o0.
302 static OptoRegPair return_value(int ideal_reg, bool is_outgoing);
303 static OptoRegPair c_return_value(int ideal_reg, bool is_outgoing);
304 RegMask _return_value_mask;
305 // Inline Cache Register
306 static OptoReg::Name inline_cache_reg();
307 static int inline_cache_reg_encode();
309 // Register for DIVI projection of divmodI
310 static RegMask divI_proj_mask();
311 // Register for MODI projection of divmodI
312 static RegMask modI_proj_mask();
314 // Register for DIVL projection of divmodL
315 static RegMask divL_proj_mask();
316 // Register for MODL projection of divmodL
317 static RegMask modL_proj_mask();
319 // Use hardware DIV instruction when it is faster than
320 // a code which use multiply for division by constant.
321 static bool use_asm_for_ldiv_by_con( jlong divisor );
323 static const RegMask method_handle_invoke_SP_save_mask();
325 // Java-Interpreter calling convention
326 // (what you use when calling between compiled-Java and Interpreted-Java
328 // Number of callee-save + always-save registers
329 // Ignores frame pointer and "special" registers
330 static int number_of_saved_registers();
332 // The Method-klass-holder may be passed in the inline_cache_reg
333 // and then expanded into the inline_cache_reg and a method_oop register
335 static OptoReg::Name interpreter_method_oop_reg();
336 static int interpreter_method_oop_reg_encode();
338 static OptoReg::Name compiler_method_oop_reg();
339 static const RegMask &compiler_method_oop_reg_mask();
340 static int compiler_method_oop_reg_encode();
342 // Interpreter's Frame Pointer Register
343 static OptoReg::Name interpreter_frame_pointer_reg();
345 // Java-Native calling convention
346 // (what you use when intercalling between Java and C++ code)
348 // Array mapping arguments to registers. Argument 0 is usually the 'this'
349 // pointer. Registers can include stack-slots and regular registers.
350 static void c_calling_convention( BasicType*, VMRegPair *, uint );
351 // Frame pointer. The frame pointer is kept at the base of the stack
352 // and so is probably the stack pointer for most machines. On Intel
353 // it is ESP. On the PowerPC it is R1. On Sparc it is SP.
354 OptoReg::Name c_frame_pointer() const;
355 static RegMask c_frame_ptr_mask;
357 // !!!!! Special stuff for building ScopeDescs
358 virtual int regnum_to_fpu_offset(int regnum);
360 // Is this branch offset small enough to be addressed by a short branch?
361 bool is_short_branch_offset(int rule, int br_size, int offset);
363 // Optional scaling for the parameter to the ClearArray/CopyArray node.
364 static const bool init_array_count_is_in_bytes;
366 // Threshold small size (in bytes) for a ClearArray/CopyArray node.
367 // Anything this size or smaller may get converted to discrete scalar stores.
368 static const int init_array_short_size;
370 // Some hardware needs 2 CMOV's for longs.
371 static const int long_cmove_cost();
373 // Some hardware have expensive CMOV for float and double.
374 static const int float_cmove_cost();
376 // Should the Matcher clone shifts on addressing modes, expecting them to
377 // be subsumed into complex addressing expressions or compute them into
378 // registers? True for Intel but false for most RISCs
379 static const bool clone_shift_expressions;
381 static bool narrow_oop_use_complex_address();
383 // Generate implicit null check for narrow oops if it can fold
384 // into address expression (x64).
385 //
386 // [R12 + narrow_oop_reg<<3 + offset] // fold into address expression
387 // NullCheck narrow_oop_reg
388 //
389 // When narrow oops can't fold into address expression (Sparc) and
390 // base is not null use decode_not_null and normal implicit null check.
391 // Note, decode_not_null node can be used here since it is referenced
392 // only on non null path but it requires special handling, see
393 // collect_null_checks():
394 //
395 // decode_not_null narrow_oop_reg, oop_reg // 'shift' and 'add base'
396 // [oop_reg + offset]
397 // NullCheck oop_reg
398 //
399 // With Zero base and when narrow oops can not fold into address
400 // expression use normal implicit null check since only shift
401 // is needed to decode narrow oop.
402 //
403 // decode narrow_oop_reg, oop_reg // only 'shift'
404 // [oop_reg + offset]
405 // NullCheck oop_reg
406 //
407 inline static bool gen_narrow_oop_implicit_null_checks() {
408 return Universe::narrow_oop_use_implicit_null_checks() &&
409 (narrow_oop_use_complex_address() ||
410 Universe::narrow_oop_base() != NULL);
411 }
413 // Is it better to copy float constants, or load them directly from memory?
414 // Intel can load a float constant from a direct address, requiring no
415 // extra registers. Most RISCs will have to materialize an address into a
416 // register first, so they may as well materialize the constant immediately.
417 static const bool rematerialize_float_constants;
419 // If CPU can load and store mis-aligned doubles directly then no fixup is
420 // needed. Else we split the double into 2 integer pieces and move it
421 // piece-by-piece. Only happens when passing doubles into C code or when
422 // calling i2c adapters as the Java calling convention forces doubles to be
423 // aligned.
424 static const bool misaligned_doubles_ok;
426 // Perform a platform dependent implicit null fixup. This is needed
427 // on windows95 to take care of some unusual register constraints.
428 void pd_implicit_null_fixup(MachNode *load, uint idx);
430 // Advertise here if the CPU requires explicit rounding operations
431 // to implement the UseStrictFP mode.
432 static const bool strict_fp_requires_explicit_rounding;
434 // Are floats conerted to double when stored to stack during deoptimization?
435 static bool float_in_double();
436 // Do ints take an entire long register or just half?
437 static const bool int_in_long;
439 // Do the processor's shift instructions only use the low 5/6 bits
440 // of the count for 32/64 bit ints? If not we need to do the masking
441 // ourselves.
442 static const bool need_masked_shift_count;
444 // This routine is run whenever a graph fails to match.
445 // If it returns, the compiler should bailout to interpreter without error.
446 // In non-product mode, SoftMatchFailure is false to detect non-canonical
447 // graphs. Print a message and exit.
448 static void soft_match_failure() {
449 if( SoftMatchFailure ) return;
450 else { fatal("SoftMatchFailure is not allowed except in product"); }
451 }
453 // Check for a following volatile memory barrier without an
454 // intervening load and thus we don't need a barrier here. We
455 // retain the Node to act as a compiler ordering barrier.
456 static bool post_store_load_barrier(const Node* mb);
459 #ifdef ASSERT
460 void dump_old2new_map(); // machine-independent to machine-dependent
462 Node* find_old_node(Node* new_node) {
463 return _new2old_map[new_node->_idx];
464 }
465 #endif
466 };
468 #endif // SHARE_VM_OPTO_MATCHER_HPP